A Human-Centered Artificial Intelligence Framework for Explainable and Trustworthy Decision Support Systems: Driver-Support System for Work-Zone Hazards

Background: Work-zone warning systems are required to identify hazardous encounters with sufficient time for human reaction, but false classifications can mask ineffective warnings.Objective: Assess explainable vehicle–worker warning models in the context of an updated physics-based pre-onset endpoint. Methods: There were two sessions in October 2018 that gave 29,273 synchronised observations, of which 7,101 were suitable for being modelled over 2 seconds. However, six models were evaluated – including weighted and unweighted XGBoost models – and compared to chronological validation and a later-session test with seven proxy events. The factors of calibration, explanation stability and sensor sensitivity and warning burden were analysed.Results: Unweighted XGBoost identified 5 out of 7 test events and generated 300 false episode warning events, while weighted XGBoost identified 4 events, and missed 3 and generated 7 false episode warning events. The mean lead times for detected events were 0.749 and 0.655 seconds, respectively. A warning was not issued for at least 1.5 seconds prior to onset for none of the seven test events, with a maximum of 1.178 seconds before onset.Opinion: Physics derived proxy onsets are not sufficient evidence for the quality of actionable warnings. The limited diversity of events, short lead times recorded and response delays that are not measured do not allow for conclusions of deployment readiness. For external safety evaluation the additional continuous recording and independent event annotations are required.